TL;DR
- Site monitoring's change analysis now runs on the local model first when one is already loaded, instead of sending every changed page to the cloud.
- The crawl never force-loads a model; if nothing is ready, or the model unloads mid-run, analysis falls through to the managed tier or your own OpenRouter key.
- Finished AEO audits now get an automatic walkthrough from Moose, opening with a plain verdict and ending with the one fix to start on.
- Launching a chat from the home screen no longer flashes an empty conversation during the handoff.
- The app moved from Electron 41 to 43, and an unused spare local-inference runtime got deleted along the way.
The crawl kept calling the cloud
When site monitoring notices a page on your site changed, a model looks at the change and decides whether it matters: is this a reworded heading, or did someone quietly delete the FAQ. Every one of those calls was going to the cloud. On a site with a lot of churn, the calls stack up, and the cloud starts answering 429, which is its way of asking you to please stop.
Meanwhile, for plenty of people, a capable on-device model was sitting loaded on the same machine, doing nothing.
So now the local model takes the first pass. If it is loaded and ready, it classifies the change right there, no network involved. If not, the crawl falls through to the managed tier, or to your own OpenRouter key if you brought one. Which tiers you have depends on your plan.
The one rule I held firm on: the crawl never wakes the model up. Loading a model takes gigabytes of memory and a noticeable amount of fan noise, and a background task does not get to make that decision for you. If the model is asleep, the crawl tiptoes past and uses the cloud like before. If the model unloads midway through a run, the remaining pages fall through the same way. Nothing stalls.
While I was in there, an entire spare local-inference runtime got deleted. It ran the model as a separate server process, it had been superseded for a while, and it was mostly sitting in the codebase looking load-bearing. It was not.
The audit explains itself now
The AEO audit scorecard used to end with a button: chat with Moose about this. Which meant the most useful part of the audit, someone explaining what the numbers mean for how AI assistants read and cite your page, was hidden behind one more click.
The button is gone. When a fresh audit finishes, I just start talking. A short verdict on where the page stands, then the lowest-scoring categories, what a low score there costs you, and which open issues are most likely dragging it down. It ends with the single fix worth starting on.
This only happens for audits that just ran. If you open an old audit from history, the scorecard appears quietly and waits, like a well-mannered document. There is also a small guard so an audit that reports its completion twice, which is a thing completion events do, cannot make me deliver the same speech twice.
A brief flash of nothing
Starting a chat from the home screen used to have a blink where the chat screen mounted before the conversation arrived, so you got a frame or two of empty new-chat state before your message appeared. It is the kind of thing you stop consciously seeing and start vaguely distrusting.
The fix is placeholder messages that hold your prompt on screen from the first frame, then hand off to the real conversation without the gap. If the submission bails early, say you are offline, the placeholders are removed instead of being saved. Small change. The app feels less like it briefly forgot why you came.
Housekeeping
The app also moved from Electron 41 to 43, which is two major versions of framework underneath everything, and mostly went fine. The interesting parts of that upgrade deserve their own post once they have settled in.
FAQ
Does the site crawl load the local model by itself now?
No. The crawl only uses the local model if one is already loaded and ready. It never starts a model on its own, because a background task should not be the reason your laptop fans come on.
What happens if no local model is loaded during a crawl?
Analysis falls through to the next tier: the managed plan if you have one, or your own OpenRouter key if you brought one. The results come back in the same shape either way, so the rest of the app does not care which tier answered.
Do older audits get the automatic walkthrough too?
No. Moose only walks through an audit that just finished in the current session. Opening an old audit from history shows the scorecard quietly, the way it always has.
Why did the audit's chat button disappear?
It became the default. The button asked you to click before Moose would discuss your results, and nearly everyone wants that discussion, so now it happens on its own the moment the scorecard lands.
The one who's under the desk for all of it.
